Family guide · 12 November 2026
Causal diagnosis and repair
Acting successfully, identifying a cause, and preserving its evidence are different tasks. This family links two technical reports; it is not a new combined manuscript or a criterion for consciousness.
| Question | Start here | Contract boundary |
|---|---|---|
| Does explicitly naming an actuator estimate a self-model improve control? | R4 finite-state pilot | A changing actuator with prescribed feedback and a short horizon. Matched explicit, compiled and external trackers tie exactly. R4 remains a separate programme entry. |
| Which cause or target can permitted noisy tests distinguish? | Causal Self-Diagnosis v0.3 — foundations report | Main oracle results fix the diagnostic target and use known repeatable observation laws. Executing a test may change future access. |
| Can repair restore health without losing the original-fault distinction? | Diagnosing While Repairing v0.2.5 | Resets change physical state and current readout ports. Its origin/restoration theorems are not automatic consequences of the fixed-state oracle. |
Control can succeed while the cause remains ambiguous
In the foundations example, routine output depends on command XOR actuator reversal XOR environmental reversal. Learning the combined reversal can be enough to compensate in the present environment, while leaving the two causes indistinguishable. After transfer to a known unreversed environment, the actuator bit matters separately. With independent fair actuator/world bits, routine evidence alone leaves it at probability one half. A trusted reference changes the experiment; merely calling an action an intervention does not guarantee attribution.
Testing must preserve the task that matters
A homotopy-preserving change can remove a needed recovery route. Conversely, retaining a route to an acceptable target can change topology. The foundations report gives three-state examples and task-relative access rules. A concrete device needs independently justified transitions, safe/target sets, costs and observation laws. Neither a topological inclusion nor a label of formal permission supplies that engineering guarantee.
Read by question
- General introduction: R4’s question and controlled null → foundations §7 self/world example → repair manuscript §§1–3.
- Noisy access: foundations §§3–6 for support, law classes and attainable terminal channels. A family of arbitrarily accurate finite procedures is different from one ongoing consistent estimator.
- Recovery: foundations §§10–11 for implementation assumptions and task preservation; then the repair manuscript’s changing-state contract.
- Applications: foundations §§12–14 for observation fidelity, anchors, motion and resource simulations; repair Appendix C for a separate proposed reset benchmark.
Limits shared across the family
Exact support is not a high-confidence subset. Known fresh repeatable kernels are not merely matching one-read marginals. Per-read depletion, hidden preparation error or physical state changes can invalidate a transfer. The motor studies are simulations, not hardware observations and not validation of the repair ports. Bayesian filtering, active testing, convex decision theory and controlled-information comparison have established antecedents. The targeted internal AI final-source review of v0.3 completed on 12 November 2026 found no blocking error under the declared assumptions. Internal AI reviews do not establish external peer review or historical priority.
The foundations PDF preserves its dated byline, open submission authorship and source-audit caveats. Historical necessity-claim challenges are not official errata. Newer repair antecedent work should not be treated as already integrated into the frozen v0.3. Acquirable Predictive Models supplies a related acquisition/prediction reading path under its own contracts.